|Publication number||US5301109 A|
|Application number||US 07/734,291|
|Publication date||5 Apr 1994|
|Filing date||17 Jul 1991|
|Priority date||11 Jun 1990|
|Publication number||07734291, 734291, US 5301109 A, US 5301109A, US-A-5301109, US5301109 A, US5301109A|
|Inventors||Thomas K. Landauer, Michael L. Littman|
|Original Assignee||Bell Communications Research, Inc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (7), Referenced by (270), Classifications (12), Legal Events (7)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This is a continuation-in-part of application Ser. No. 07/536,029, filed Jun. 11, 1990 now abandoned.
This invention relates generally to computer-based information retrieval and, in particular, to user accessibility to and display of textual material stored in computer files utilizing a request in one language to retrieve documents in other languages related to the request.
In the field of information retrieval, a long-standing objective has been the development of an automated procedure by which documents in one language could be effectively accessed by requests in another language without needing to translate either the documents or the requests. Among other things, such a capability would allow users to determine what documents were available in languages that the users could not read before incurring the expense and delay of translation.
One technique representative of some previously proposed procedures, disclosed in an article entitled "Automatic Processing of Foreign Language Documents," was published by G. Salton in 1970 in the Journal of American Society for Information Sciences. Salton reported experimenting with a method for automatic retrieval of documents in one language in response to queries in another using a vector representation and search technique in conjunction with a manually created dual-language thesaurus. The results for test samples of abstracts and queries were promising. However, creating an adequate multi-language thesaurus is difficult and requires considerable intellectual labor. Moreover, a traditional thesaurus necessarily imposes a discrete and rather restricted model of the languages in question and of their relation to one another.
U.S. Pat. No. 4,839,853, issued to one of the present co-inventors and assigned to the same assignee as is the present invention, utilizes the Latent Semantic Indexing (LSI) approach to model the underlying correlational structure of the distribution of terms in documents. Instead of representing documents and queries directly as sets of words, the LSI technique represents them as parameters in such a way that dependencies between words and between documents are taken into account. For example, if two terms are used in exactly the same contexts, that is, have identical distribution across a target collection of documents, LSI is designed to treat them not as two independent indexing entries but as two instances of an abstract indexing variable with the same vector value. Lesser and more indirect relations between terms and between documents are represented in an appropriate analogous fashion.
In the implementation of LSI as set forth in the above-identified patent, the modeling is accomplished by approximating the original term-by-document matrix by the product of three lower rank matrices of orthogonal derived indexing variables. The first matrix represents terms as values on a smaller set of independent "basis" vectors; the second matrix contains scaling coefficients; and the third matrix represents documents as values on the smaller set of basis vectors. The method can be interpreted geometrically as a means by which each document and each term is assigned to a point in a hyperspace. The mathematics and implementation of the method construct a derived space in which terms, documents, and queries can all be represented in the hyperspace. The mathematical procedure employed is singular value decomposition (SVD), which is closely related to factor analysis and eigenvalue decomposition.
The retrieval process is the same as in standard vector methods, e.g. using document-query cosines as the similarity measure. Various preprocessing steps, such as term weighting, may also be done in standard ways. The principal difference between LSI and previous vector models as represented by the work of Salton is that the vectors are constructed in a space with many fewer dimensions than the number of original terms, and that these dimensions are the subset of linearly independent basis vectors by which the original term-by-document matrix can be best approximated in a least squares sense. The number of dimensions retained has been determined empirically; optimal retrieval performance has usually been obtained with about 100 dimensions for collections of many hundreds to several thousands of documents.
The dimension reduction step of LSI has the advantageous property that small sources of variability in term usage are dropped and only the most important sources kept. Among other things, this can cause synonyms or near synonyms to be collapsed into similar vector representations, with the result that queries can retrieve similar documents even though they share no terms. This cannot happen in the usual raw term vector representation, necessitating manually constructed thesauri with their attendant problems.
The LSI method has previously been applied only within a single language, and there has been no teaching or suggestion in the art regarding the application of LSI to multi-language information retrieval.
These shortcomings as well as other deficiencies and limitations of conventional information retrieval techniques are obviated, in accordance with the present invention, by constructing a multi-language semantic space. This is effected automatically, without the need for a thesaurus, by modeling the usage of terms in documents using an expanded latent semantic indexing framework. In the broad aspect of the method, an initial set of documents, from a usually larger set of documents, is translated into the number of languages under consideration and the documents, including all translations, are stored in a computer information file; this produces a set of multiple (dual in one special but significant case) language documents. This set of multi-lingual documents is used to "train" an automatic multi-lingual indexing system by processing a joint term-by-document matrix of data. The joint matrix is formed by including the terms used in all the translations, and each document is allocated a single vector in the matrix no matter how many languages are treated by the methodology. After training, i.e., application of the singular value decomposition, the system can index any new document or query that is presented to it according to a set of derived abstract indexing variables that are language-independent.
The organization and operation of this invention will be better understood from a consideration of the detailed description, which follows, when taken in conjunction with the accompanying drawing.
FIG. 1 is a plot of the "term" coordinates and the "document" coordinates based on a two-dimensional singular value decomposition of an original "term-by-document" matrix in a single language;
FIG. 2 shows the location of the training documents in the data object space for an example reduced to two dimensions in a dual language example; and
FIG. 3 is a flow diagram depicting the processing which generates the "term" and "document" matrices using singular value decomposition as well as the processing of a user's query.
Before discussing the principles and operational characteristics of this invention in detail, it is helpful to present a motivating example of latent semantic indexing for a single language case, namely, English. This also aids in introducing terminology utilized later in the discussion.
The contents of Table 1 are used to illustrate how semantic structure analysis works and to point out the differences between this method and conventional keyword matching.
c1: Human machine interface for Lab ABC computer applications
c2: A survey of user opinion of computer system response time
c3: The EPS user interface management system
c4: Systems and human systems engineering testing of EPS-2
c5: Relation of user-perceived response time to error measurement
m1: The generation of random, binary, unordered trees
m2: The intersection graph of paths in trees
m3: Graph minors IV: Widths of trees and well-quasi-ordering
m4: Graph minors: A survey
In this example, a file of text objects consists of nine titles of technical documents with titles c1-c5 concerned with human/computer interaction and titles m1-m4 concerned with mathematical graph theory. In Table 1, words occuring in more than one title are italicized. Using conventional keyword retrieval, if a user requested papers dealing with "human computer interaction," titles c1, c2, and c4 would be returned, since these titles contain at least one keyword from the user request. However, c3 and c5, while related to the query, would not be returned since they share no words in common with the request. It is now shown how latent semantic structure analysis treats this request to return titles c3 and c5.
Table 2 depicts the "term-by-document" matrix for the 9 technical document titles. Each cell entry, (i,j), is the frequency of occurrence of term i in document j. This basic term-by-document matrix or a mathematical transformation thereof is used as input to the statistical procedure described below.
TABLE 2______________________________________ DOCUMENTSTERMS c1 c2 c3 c4 c5 m1 m2 m3 m4______________________________________ 1. human 1 0 0 1 0 0 0 0 0 2. interface 1 0 1 0 0 0 0 0 0 3. computer 1 1 0 0 0 0 0 0 0 4. user 0 1 1 0 1 0 0 0 0 5. system 0 1 1 2 0 0 0 0 0 6. response 0 1 0 0 1 0 0 0 0 7. time 0 1 0 0 1 0 0 0 0 8. EPS 0 0 1 1 0 0 0 0 0 9. survey 0 1 0 0 0 0 0 0 110. tree 0 0 0 0 0 1 1 1 011. graph 0 0 0 0 0 0 1 1 112. minor 0 0 0 0 0 0 0 1 1______________________________________
For this example the documents and terms have been carefully selected to yield a good approximation in just two dimensions for expository purposes. FIG. 1 is a two-dimensional graphical representation of the two largest dimensions resulting from the mathematical process, singular value decomposition. Both document titles and the terms used in them are placed into the same space. Terms are shown as circles and labeled by number. Document titles are represented by squares with the numbers of constituent terms indicated parenthetically. The angle between two objects (terms or documents) describe their computed similarity. In this representation, the two types of documents form two distinct groups: all the mathematical graph theory titles occupy the same region in space (basically along Dimension 1 of FIG. 1) whereas a quite distinct group is formed for human/computer interaction titles (essentially along Dimension 2 of FIG. 1).
To respond to a user query about "human computer interaction," the query is first folded into this two-dimensional space using those query terms that occur in the space (namely, "human" and "computer"). The query vector is located in the direction of the weighted average of these constituent terms, and is denoted by a directional arrow labeled "Q" in FIG. 1. A measure of closeness or similarity is the angle between the query vector and any given term or document vector. In FIG. 1 the cosine between the query vector and each c1-c5 titles is greater than 0.90; the angle corresponding to the cosine value of 0.90 with the query is shown by the dashed lines in FIG. 1. With this technique, documents c3 and c5 would be returned as matches to the user query, even though they share no common terms with the query. This is because the latent semantic structure (represented in FIG. 1) fits the overall pattern of term usage across documents.
To obtain the data to plot FIG. 1, the "term-by-document" matrix of Table 2 is decomposed using singular value decomposition (SVD). A reduced SVD is employed to approximate the original matrix in terms of a much smaller number of orthogonal dimensions. The reduced dimensional matrices are used for retrieval; these describe major associational structures in the term-document matrix but ignore small variations in word usage. The number of dimensions to represent adequately a particular domain is largely an empirical matter. If the number of dimensions is too large, random noise or variations in word usage will be modeled. If the number of dimensions is too small, significant semantic content will remain uncaptured. For diverse information sources, 100 or more dimensions may be needed.
To illustrate the decomposition technique, the term-by-document matrix, denoted Y, is decomposed into three other matrices, namely, the term matrix (TERM), the document matrix (DOCUMENT), and a diagonal matrix of singular values (DIAGONAL), as follows:
Yt,d =TERMt,k DIAGONALk,k DOCUMENTT k,d
where Y is the original t-by-d matrix, TERM is the t-by-k matrix that has unit-length orthogonal columns, DOCUMENTT is the transpose of the d-by-k DOCUMENT matrix with unit-length orthogonal columns, and DIAGONAL is the k-by-k diagonal matrix of singular values typically ordered by magnitude.
The dimensionality of the solution, denoted k, is the rank of the t-by-d matrix, that is, k≦min(t,d). Tables 3, 4 and 5 below show the TERM and DOCUMENT matrices and the diagonal elements of the DIAGONAL matrix, respectively, as found via SVD.
TABLE 3__________________________________________________________________________TERM MATRIX (12 terms by 9 dimensions)__________________________________________________________________________human0.22 -0.11 0.29 -0.41 -0.11 -0.34 -.52 -0.06 -0.41interface0.20 -0.07 0.14 -0.55 0.28 0.50 -0.07 -0.01 -0.11computer0.24 0.04 -0.16 -0.59 -0.11 -0.25 -0.30 0.06 0.49user 0.40 0.06 -0.34 0.10 0.33 0.38 0.00 0.00 0.01system0.64 -0.17 0.36 0.33 -0.16 -0.21 -0.16 0.03 0.27response0.26 0.11 -0.42 0.07 0.08 -0.17 0.28 -0.02 -0.05time 0.26 0.11 -0.42 0.07 0.08 -0.17 0.28 -0.02 -0.05EPS 0.30 -0.14 0.33 0.19 0.11 0.27 0.03 -0.02 -0.16survey0.20 0.27 -0.18 -0.03 -0.54 0.08 -0.47 -0.04 -0.58tree 0.01 0.49 0.23 0.02 0.59 -0.39 -0.29 0.25 -0.22graph0.04 0.62 0.22 0.00 -0.07 0.11 0.16 -0.68 0.23minor0.03 0.45 0.14 -0.01 -0.30 0.28 0.34 0.68 0.18__________________________________________________________________________
TABLE 4__________________________________________________________________________DOCUMENT MATRIX (9 documents by 9 dimensions)__________________________________________________________________________c1 0.20-0.06 0.11 -0.95 0.04 -0.08 0.18 -0.01 -0.06c2 0.600.16 -0.50 -0.03 -0.21 -0.02 -0.43 0.05 0.24c3 0.46-0.13 0.21 0.04 0.38 0.07 -0.24 0.01 0.02c4 0.54-0.23 0.57 0.27 -0.20 -0.04 0.26 -0.02 -0.08c5 0.280.11 -0.50 0.15 0.33 0.03 0.67 -0.06 -0.26m1 0.000.19 0.10 0.02 0.39 -0.30 -0.34 0.45 -0.62m2 0.010.44 0.19 0.02 0.35 -0.21 -0.15 -0.76 0.02m3 0.020.62 0.25 0.01 0.15 0.00 0.25 0.45 0.52m4 0.080.53 0.08 -0.02 -0.60 0.36 0.04 -0.07 -0.45__________________________________________________________________________
TABLE 5______________________________________DIAGONAL (9 singular values)______________________________________3.34 2.54 2.35 1.64 1.50 1.31 0.84 0.56 0.36______________________________________
As alluded to earlier, data to plot FIG. 1 was obtained by presuming that two dimensions are sufficient to capture the major associational structure of the t-by-d matrix, that is, k is set to two in the expression for Yt,d, yielding an approximation of the original matrix. Only the first two columns of the TERM and DOCUMENT matrices are considered with the remaining columns being ignored. Thus, the term data point corresponding to "human" in FIG. 1 is plotted with coordinates (0.22,-0.11), which are extracted from the first row and the two left-most columns of the TERM matrix. Similarly, the document data point corresponding to title m1 has coordinates (0.00,0.19), coming from row six and the two left-most columns of the DOCUMENT matrix. Finally, the Q vector is located from the weighted average of the terms "human" and "computer" appearing in the query. A method to compute the weighted average will be presented below.
It is now elucidating to describe in somewhat more detail the mathematical model underlying the latent structure, singular value decomposition technique.
Any rectangular matrix Y of t rows and d columns, for example, a t-by-d matrix of terms and documents, can be decomposed into a product of three other matrices:
Y=To So DT o, (1)
such that To and Do have unit-length orthogonal columns (i.e. To T To =I; Do T Do =I) and So is diagonal. This is called the singular value decomposition (SVD) of Y. (A procedure for SVD is described in the text Numerical Recipes, by Press, Flannery, Teukolsky and Vetterling, 1986, Cambridge University Press, Cambridge, England). To and Do are the matrices of left and right singular vectors and So is the diagonal matrix of singular values. By convention, the diagonal elements of So are ordered in decreasing magnitude.
With SVD, it is possible to devise a simple strategy for an optimal approximation to Y using smaller matrices. The k largest singular values and their associated columns in To and Do may be kept and the remaining entries set to zero. The product of the resulting matrices is a matrix YR which is approximately equal to Y, and is of rank k. The new matrix YR is the matrix of rank k which is the closest in the least squares sense to Y. Since zeros were introduced into So, the representation of So can be simplified by deleting the rows and columns having these zeros to obtain a new diagonal matrix S, and then deleting the corresponding columns of To and Do to define new matrices T and D, respectively. The result is a reduced model such that
YR =TSDT. (2)
The value of k is chosen for each application; it is generally such that k≧100 for collections of 1000-3000 data objects.
For discussion purposes, it is useful to interpret the SVD geometrically. The rows of the reduced matrices T and D may be taken as vectors representing the terms and documents, respectively, in a k-dimensional space. With appropriate rescaling of the axes, by quantities related to the associated diagonal values of S, dot products between points in the space can be used to access and compare objects. (A simplified approach which did not involve rescaling was used to plot the data of FIG. 1, but this was strictly for expository purposes.) These techniques are now discussed.
There are basically three types of comparisons of interest: (i) those comparing two terms; (ii) those comparing two documents or text objects; and (iii) those comparing a term and a document or text object. As used throughout, the notion of a text object or data object is general whereas a document is a specific instance of a text object or data object. Also, text or data objects are stored in the computer system in files.
Two Terms: In the data, the dot product between two row vectors of YR tells the extent to which two terms have a similar pattern of occurrence across the set of documents. The matrix YR YT R is the square symmetric matrix approximation containing all the term-by-term dot products. Using equation (2),
YR YT R =(TSDT)(TSDT)T =TS2 TT =(TS)(TS)T. (3)
This means that the dot product between the i-th row and j-th row of YR can be obtained by calculating the dot product between the i-th and j-th rows of the TS matrix. That is, considering the rows of TS as vectors representing the terms, dot products between these vectors give the comparison between the terms. The relation between taking the rows of T as vectors and those of TS as vectors is simple since S is a diagonal matrix; each vector element has been stretched or shrunk by the corresponding element of S.
Two Documents: In this case, the dot product is between two column vectors of Y. The document-to-document dot product is approximated by
YT R YR =(TSDT)T (TSDT)=DS2 DT =(DS)(DS)T. (4)
Thus the rows of the DS matrix are taken as vectors representing the documents, and the comparison is via the dot product between the rows of the DS matrix.
Term and Document: This comparison is somewhat different. Instead of trying to estimate the dot product between rows or between columns of Y, the fundamental comparison between a term and a document is the value of an individual cell in Y. The approximation of Y is simply equation (2), i.e., YR =TSDT. The i,j cell of YR may therefore be obtained by taking the dot product between the i-th row of the matrix TS1/2 and the j-th row of the matrix DS1/2. While the "within" (term or document) comparisons involved using rows of TS and DS as vectors, the "between" comparision requires TS1/2 and DS1/2 for coordinates. Thus it is not possible to make a single configuration of points in a space that will allow both "between" and "within" comparisons. They will be similar, however, differing only by a stretching or shrinking of the dimensional elements by a factor S1/2.
The previous results show how it is possible to compute comparisons between the various objects associated with the rows or columns of Y. It is very important in information retrieval applications to compute similar comparison quantities for objects such as queries that do not appear explicitly in Y. This is particularly important for the cross-language case considered in accordance with the present invention. For example, it is necessary to be able to take a completely novel query, find a location in the k-dimensional latent semantic space for it, and then evaluate its cosine with respect to terms or objects in the space. Another example would be trying, after-the-fact, to find representations for documents that did not appear in the original space. The new objects for both these examples are equivalent to objects in the matrix Y in that they may be represented as vectors of terms. For this reason they are called pseudo-documents specifically or pseudo-objects generically. In order to compare pseudo-documents to other documents, the starting point is defining a pseudo-document vector, designated Yq. Then a representation Dq is derived such that Dq can be used just like a row of D in the comparison relationships described in the foregoing sections. One criterion for such a derivation is that the insertion of a real document Yi should give Di when the model is ideal (i.e., Y=YR). With this constraint,
Yq =TSDq T
or, since TT T equals the identity matrix,
Dq T =S-1 TT Yq
Dq =YT q TS-1. (5)
Thus, with appropriate rescaling of the axes, this amounts to placing the pseudo-object at the vector sum of its corresponding term points. The Dq may be used like any row of D and, appropriately scaled by S or S1/2, can be used like a usual document vector for making "within" and "between" comparisons. [It is to be noted that if the measure of similarity to be used in comparing the query against all the documents is one in which only the angle between the vectors is important (such as the cosine), there is no difference for comparison purposes between placing the query at the vector average or the vector sum of its terms since the average and sum differ only in magnitude.]
For the query example above ("human computer interaction"), Yq =[1010 . . . ]T, so for the simplified two-dimensional representation, ##EQU1## or, finally,
Thus, Dq represents the location of the query in the document space and is basically the weighted average of the terms appearing in the query.
To extend the principles of LSI to cross-language retrieval, a document set comprising all documents of interest, in the languages to be searched, is formed. A subset of the documents, called the "training set," is selected; the "training set" is composed of documents for which translations exist in all the languages (two or more). The so-called "joint" term-by-document matrix of this set is composed from the addition of the terms in their renditions in all the languages. This joint matrix differs from the single-language LSI matrix in that each column, which represents a single multi-language document, is the combination of terms from the two (or more) languages coalesced into just a single column vector. As with the single-language technique, the joint matrix is then analyzed by singular value decomposition. The resulting representation defines vectors for the training-set terms and documents in the languages under consideration. Once the training analysis has been completed, other single-language documents can be "folded in" as pseudo-documents on the basis of terms from any one of the original languages alone. Most importantly, a user query is treated as such a new document.
In the derived indexing space there is a point representing each term in the training set. A new single-language document is assigned a point in the same space by putting it at an appropriate average of the location of all the terms it contains. For cross-language retrieval, the same number or greater of dimensions are kept as would be required to represent the collection in a single language. As outlined above, full or partial equivalence (in the sense that one term will have the same or similar effect in referencing documents as another) is induced between any two or more terms approximately to the extent that their pattern of use, or the overall pattern of association between other terms with which they co-occur, is similar across documents in the training set. Equivalent or nearly equivalent terms in different languages would, of course, be expected to be distributed in nearly the same way in a set of documents and their translations. Thus, the location of two or more equivalent terms in different languages should be almost the same in the resulting representation. Consequently, a document folded in by terms in one language is retrieved by a query containing the appropriate set of words in another language.
A simple example may aid in understanding the general procedure. For this example, a training set of "documents" is composed of four titles, each of which is stated in both English and French.
Training Doc. T1. Effect of falling oil prices on small companies. Les consequences de la chute des prix du petrole pour les petites compagnies.
Training Doc. T2. Low oil prices--Effect on Calgary. La baisse des prix petroliers--Les consequences pour les citoyens de Calgary.
Training Doc. T3. Canadian nuclear power stations--Safety precautions. Les reacteurs nucleaires canadiens--Les precautions prises pour en assurer la securite.
Training Doc. T4. Safety standards for nuclear power plants--Swedish call for international conference. Les normes de securite en matiere de centrales nucleaires--L'appel de la Suede en faveur d'une conference internationale.
First the 55 (20 English-only, 32 French-only, and 3 both) joint term-by-four document training matrix formed from these "documents" is constructed, as partially depicted in TABLE 6; this table shows the first six English-only words, the three words shared by both languages, and the the last three French-only words. It is this joint matrix that will be decomposed by SVD.
TABLE 6______________________________________ DOCUMENTSTERMS T1(e1, f1) T2(e2, f2) T3(e3, f3) T4(e4, f4)______________________________________effect 1 1 0 0of 1 0 0 0falling 1 0 0 0oil 1 1 0 0prices 1 1 0 0on 1 1 0 0Calgary 0 2 0 0precautions 0 0 2 0conference 0 0 0 2d 0 0 0 1une 0 0 0 1internationale 0 0 0 1______________________________________
As is apparent from the joint term-by-document training matrix of Table 6, each document is composed of all the terms in both French and English, i.e. the addition of terms from each document including its translation(s). For instance, since the term precautions appears as the same term in both the English and French versions, there is an entry of "2" under title T3 in the precautions row. As suggested by the foregoing illustrative example, the general procedure for formulating the joint term-by-document matrix for the multi-language case is as follows:
(1) for each document in the training set written in an original language, translate this document into all the other languages. (In the above example, each of the four training documents is in English, which is considered the original language, and each is translated to one other language, namely, French);
(2) each original document plus all of the other translations of each original document are parsed to extract distinct terms composing the multi-language documents. These terms define a database designated the lexicon database, and this database is stored in a memory of a computer. The lexicon database is used in constructing the general joint term-by-document matrix as presented below. (In the above example, the first document contained eight (8) distinct English terms and twelve (12) distinct French terms--"les" is repeated; the second document contains only two (2) more distinct English terms not contained in the first English document, namely, "low" and "Calgary". The terms "oil", "prices", "effect", and "on" are already in the lexicon database as a result of parsing the first English document. Continued parsing in this manner results in the fifty-five (55) distinct terms presented above, namely, 20 English-only, 32 French-only and 3 terms common to both languages.)
(3) the distinct terms from the lexicon database are then treated as being listed in a column, such as the TERMS column in TABLE 6, as an aid in preparing the joint term-by-document matrix; this column contains t rows. Each training document, composed of both the original as well as all translations, is assigned one column in the joint matrix; if there are d training documents, then there are d columns. Any (i,j) cell in the joint term-by-document matrix, that is, the intersection of the ith "term" row with the jth "document" column contains a tabulation of the frequency of occurrence of the term in the ith row with the document assigned to the jth column. (In the example, training document T2 is shown to have a tabulation of 1 in the row with the term "effect" since it appears only once in the coalesced or merged English and French versions of the document. In contrast, there is an entry of 2 in the row with the term "Calgary" since it appears twice in the documents of T2, namely, once in the English document and once in the French document.)
It is important to understand that it is not necessary to use all available documents to compose the training set. One useful test for the number of documents to include in the training set is the satisfactory retrieval of a document written in one language as determined by inputing the terms of the document as a query in another language. One illustrative test for the sufficiency of the training set will be presented below after the joint term-by-document matrix is decomposed. Also, it is important to realize that some retrieval situations will not require assigning all terms obtained during the parsing step to the lexicon database. A test of what terms to assign to the database is again the satisfactory retrieval of a document written in one language as determined by inputing the terms of the document as a query in another language.
By way of terminology, the generalization of a `document` is called a `data object` so as to cover applications such as graphics-type information as well as text. Moreover, the coalesced version of all translations of a data object as well as the original data object is called a merged data object.
The results of the decomposition are shown in Tables 7, 8, and 9 (which are similar to Tables 3, 4, and 5) for two dimensions.
TABLE 7______________________________________TERM MATRIX (55 terms by 2 dimensions)______________________________________effect 0.0039 -0.1962of 0.0042 -0.2550falling 0.0042 -0.2550oil 0.0039 -0.1962prices 0.0039 -0.1962on 0.0039 -0.1962Calgary 0.0056 -0.2178precautions 0.0451 -0.0036conference 0.3299 0.0124d 0.2081 0.0078une 0.2081 0.0078internationale 0.2081 0.0078______________________________________
TABLE 8______________________________________DOCUMENT MATRIX (4 documents by 2 dimensions)______________________________________T1 0.0200 -0.8799T2 0.0169 -0.4743T3 0.1355 -0.0079T4 0.9904 0.0269______________________________________
TABLE 9______________________________________DIAGONAL (2 singular values)______________________________________ 3.2986 2.3920______________________________________
FIG. 2 shows the location of the four training documents in this space. (Since the angle of the coordinates representative of each document is the important parameter for search purposes and the absolute magnitude of the coordinates of each document is relatively unimportant for search purposes, the magnitude of each document has been normalized to unit magnitude for clarity of presentation).
Next, all single-language documents are folded into the space derived from the training set. Each remaining document is folded into the resulting space separately in its English and French versions, i.e. using only English terms and then only French terms in the pseudo-document representation of equation (5): for instance,
New Doc Ne. Ontario--Premier's rejection of further nuclear power plants.
(Absolute coordinates of 0.0695,-0.0708)
New Doc Nf. L'ontario--le refus du premier ministre de favoriser la construction d'autres centrales nucleaires.
(Absolute coordinates of 0.1533,-0.0775)
As shown, the English-only and French-only versions, Ne and Nf, end up close ("similar") to one another and well separated from the other text items in the space. In fact, for a search angle of approximately plus/minus 26 degrees (cosine of 0.90), each document falls within the angle of similarity of the other document. The degree of similarity or closeness of corresponding documents folded into the semantic space after training is used as a test for the sufficiency of the set of data objects selected to train the semantic space. For instance, after training, if a set of documents like Ne and Nf does not fall within a preselected angle of similarity, then it may be necessary to re-train the semantic space in order to meet the prescribed retrieval criterion/criteria--for the illustrative case, a single criterion is falling within the angle of search. Typically, paragraphs of 50 words or more from 500 or more multi-language documents are suitable to train the semantic space.
Of course, it is possible to have considered the folded in documents from another viewpoint, namely: one might have been an untranslated document entered in one language into the indexing space for later reference, and the other an untranslated query in another language. Indeed, the latter viewpoint is the intended manner of use, and a query in one language would locate the translated version of the document.
Having defined a cross-language indexing space on the basis of an initial sample of multi-lingual documents, new documents would be entered without translation, using words from their original language only. Similarly, queries would be entered in whatever language they are posed by the user. The cross-language indesing space will make it possible to match a query in any language with a document in any language.
For the trans-language case discussed here, if it were assumed that the only difference between the two language versions was the orthography of individual words, i.e. that a word-for-word correspondence between languages existed for the collection in question, then it would be clear that the same number of dimensions would be optimal for the joint representation as for any one of the languages alone. The dimension reduction would thus implicitly include a factor of l reduction for the number l of languages involved. The expected result would be that differences between languages would be perfectly collapsed in the resulting representation; a term in any language would map to the identical vector value to any of its exact translations. Of course, languages do not translate perfectly by one to one word substitution, so the true optimum number of dimensions for a joint representation might be somewhat larger (or perhaps smaller) than for each single language, in order, for example, to capture important structure special to each language separately. It has been found that the k for the single language case has been provided acceptable performance for the multi-language case.
The equations given above are independent of any preprocessing or reweighting of the rows or columns of Y. Such preprocessing might be used to prevent documents of different overall length from having differential effect on the model, or be used to impose preconceptions of which terms are more important.
The foundation principles presented in the foregoing sections are now described in a process flow manner by way of teaching an illustrative embodiment in accordance with the present invention.
The technique for processing the documents and their corresponding translations is shown in block diagram form in FIG. 3. The first processing phase, as represented by blocks 100-150, is that of training the system. Initially, as represented by block 100, the set of training documents is selected.
The next processing activity of the training phase, as illustrated by processing block 110, is that of preprocessing the training set.
The next step to the processing is represented by block 120 in FIG. 3. Based upon the earlier text preprocessing, a system lexicon of terms is created. Such a processing step is accomplished, for example, by parsing the data objects to obtain selected terms, such as all nouns, verbs, adjectives, and so forth.
From the list of lexicon terms, the Joint Term-by-Document matrix is created, as depicted by processing block 130 in FIG. 3.
The next step performed in the training phase is the singular value decomposition on the Joint Term-by-Document matrix, as depicted by processing block 140. This analysis is only effected once (or each time there is a significant update in the storage files).
The final processing step in the training phase is the generation of the term database, as depicted by block 150.
The fold-in phase is represented by blocks 160-190. All single-language documents are preprocessed in the same manner as block 110, that is, terms are isolated and reduced to lowercase; this is depicted by processing block 160.
Next, as depicted by processing block 170, terms from each single language document are located in the terms-database.
Following this processing, processing block 180 is invoked to compute the spatial coordinates of each single language document as per equation (5).
Finally, the total document database, including the training documents, is constructed via processing of block 190.
The user query processing activity is depicted on the right-most side of FIG. 3. The first step, as represented by processing block 200, is to preprocess the query in the same way as the original documents.
As then depicted by block 210, for each query term also contained in the system lexicon, the k-dimensional vector is constructed.
Processing block 220 depicts processing of the query vector using equation (5).
The next step in the query processing is depicted by processing block 230. In order that the best matching document is located, the query vector is compared to all documents in the space. The similarity metric used is the cosine or dot product between the query vector and the document vectors. (The cosine metric is similar to a dot product measure except that it ignores the magnitude of the vectors and simply uses the angle between the vectors being compared.) A cosine of 1.0 would indicate that the query vector and the document vector were on top of one another in the space. Typically an angle having a cosine of at least 0.90 is considered a good match.
The cosines are sorted, as depicted by processing block 240. Finally, as shown in block 250, the documents within the angle defined by the desired cosine paramenter are stored for later recall by the user with the option of displaying a subset of the requested documents.
It is to be further understood that the methodology described herein is not limited to the specific forms disclosed by way of illustration, but may assume other embodiments limited only by the scope of the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4270182 *||30 Dec 1974||26 May 1981||Asija Satya P||Automated information input, storage, and retrieval system|
|US4593356 *||16 Jul 1981||3 Jun 1986||Sharp Kabushiki Kaisha||Electronic translator for specifying a sentence with at least one key word|
|US4654798 *||14 Aug 1984||31 Mar 1987||Mitsubishi Denki Kabushiki Kaisha||System of simultaneous translation into a plurality of languages with sentence forming capabilities|
|US4839853 *||15 Sep 1988||13 Jun 1989||Bell Communications Research, Inc.||Computer information retrieval using latent semantic structure|
|US4864503 *||5 Feb 1987||5 Sep 1989||Toltran, Ltd.||Method of using a created international language as an intermediate pathway in translation between two national languages|
|US4916614 *||23 Nov 1987||10 Apr 1990||Hitachi, Ltd.||Sentence translator using a thesaurus and a concept-organized co- occurrence dictionary to select from a plurality of equivalent target words|
|US5128865 *||2 Mar 1990||7 Jul 1992||Bso/Buro Voor Systeemontwikkeling B.V.||Method for determining the semantic relatedness of lexical items in a text|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US5404514 *||13 Sep 1993||4 Apr 1995||Kageneck; Karl-Erbo G.||Method of indexing and retrieval of electronically-stored documents|
|US5418948 *||8 Sep 1993||23 May 1995||West Publishing Company||Concept matching of natural language queries with a database of document concepts|
|US5440481 *||28 Oct 1992||8 Aug 1995||The United States Of America As Represented By The Secretary Of The Navy||System and method for database tomography|
|US5442781 *||24 Nov 1992||15 Aug 1995||Nec Corporation||System for generating a search formula by accessing search terms on the basis of a training set of pertinent and non-pertinent objects|
|US5459861 *||22 Mar 1993||17 Oct 1995||Atr Auditory And Visual Perception Research Laboratories||Evaluation support method for retrieved data|
|US5488725 *||30 Mar 1993||30 Jan 1996||West Publishing Company||System of document representation retrieval by successive iterated probability sampling|
|US5544352 *||14 Jun 1993||6 Aug 1996||Libertech, Inc.||Method and apparatus for indexing, searching and displaying data|
|US5630131 *||14 Nov 1994||13 May 1997||Object Technology Licensing Corp.||Method and apparatus for importing and exporting archive files for a graphical user interface|
|US5642502 *||6 Dec 1994||24 Jun 1997||University Of Central Florida||Method and system for searching for relevant documents from a text database collection, using statistical ranking, relevancy feedback and small pieces of text|
|US5649193 *||17 Nov 1995||15 Jul 1997||Kabushiki Kaisha Toshiba||Document detection system using detection result presentation for facilitating user's comprehension|
|US5652884 *||14 Nov 1994||29 Jul 1997||Object Technology Licensing Corp.||Method and apparatus for dynamic update of an existing object in an object editor|
|US5659766 *||16 Sep 1994||19 Aug 1997||Xerox Corporation||Method and apparatus for inferring the topical content of a document based upon its lexical content without supervision|
|US5687364 *||16 Sep 1994||11 Nov 1997||Xerox Corporation||Method for learning to infer the topical content of documents based upon their lexical content|
|US5713016 *||5 Sep 1995||27 Jan 1998||Electronic Data Systems Corporation||Process and system for determining relevance|
|US5715445 *||7 Jun 1995||3 Feb 1998||Wolfe; Mark A.||Document retrieval system employing a preloading procedure|
|US5724571 *||7 Jul 1995||3 Mar 1998||Sun Microsystems, Inc.||Method and apparatus for generating query responses in a computer-based document retrieval system|
|US5727195 *||14 Apr 1995||10 Mar 1998||Fuji Xerox Co., Ltd.||Document retrieving object instructing apparatus|
|US5745893 *||30 Nov 1995||28 Apr 1998||Electronic Data Systems Corporation||Process and system for arrangement of documents|
|US5754938 *||31 Oct 1995||19 May 1998||Herz; Frederick S. M.||Pseudonymous server for system for customized electronic identification of desirable objects|
|US5754939 *||31 Oct 1995||19 May 1998||Herz; Frederick S. M.||System for generation of user profiles for a system for customized electronic identification of desirable objects|
|US5778362 *||21 Jun 1996||7 Jul 1998||Kdl Technologies Limted||Method and system for revealing information structures in collections of data items|
|US5787424 *||30 Nov 1995||28 Jul 1998||Electronic Data Systems Corporation||Process and system for recursive document retrieval|
|US5835087 *||31 Oct 1995||10 Nov 1998||Herz; Frederick S. M.||System for generation of object profiles for a system for customized electronic identification of desirable objects|
|US5870770 *||28 Jan 1998||9 Feb 1999||Wolfe; Mark A.||Document research system and method for displaying citing documents|
|US5893092 *||23 Jun 1997||6 Apr 1999||University Of Central Florida||Relevancy ranking using statistical ranking, semantics, relevancy feedback and small pieces of text|
|US5963940 *||14 Aug 1996||5 Oct 1999||Syracuse University||Natural language information retrieval system and method|
|US5987446 *||12 Nov 1996||16 Nov 1999||U.S. West, Inc.||Searching large collections of text using multiple search engines concurrently|
|US6006221 *||14 Aug 1996||21 Dec 1999||Syracuse University||Multilingual document retrieval system and method using semantic vector matching|
|US6026388 *||14 Aug 1996||15 Feb 2000||Textwise, Llc||User interface and other enhancements for natural language information retrieval system and method|
|US6029195 *||5 Dec 1997||22 Feb 2000||Herz; Frederick S. M.||System for customized electronic identification of desirable objects|
|US6055528 *||25 Jul 1997||25 Apr 2000||Claritech Corporation||Method for cross-linguistic document retrieval|
|US6076088 *||6 Feb 1997||13 Jun 2000||Paik; Woojin||Information extraction system and method using concept relation concept (CRC) triples|
|US6088692 *||5 Apr 1999||11 Jul 2000||University Of Central Florida||Natural language method and system for searching for and ranking relevant documents from a computer database|
|US6101491 *||31 Mar 1997||8 Aug 2000||Sun Microsystems, Inc.||Method and apparatus for distributed indexing and retrieval|
|US6182063||31 Mar 1997||30 Jan 2001||Sun Microsystems, Inc.||Method and apparatus for cascaded indexing and retrieval|
|US6219632||20 Nov 1997||17 Apr 2001||International Business Machines Corporation||System for the facilitation of supporting multiple concurrent languages through the use of semantic knowledge representation|
|US6256629 *||25 Nov 1998||3 Jul 2001||Lucent Technologies Inc.||Method and apparatus for measuring the degree of polysemy in polysemous words|
|US6263329 *||3 Sep 1999||17 Jul 2001||Claritech||Method and apparatus for cross-linguistic database retrieval|
|US6263335||29 Mar 1999||17 Jul 2001||Textwise Llc||Information extraction system and method using concept-relation-concept (CRC) triples|
|US6301576||31 Dec 1999||9 Oct 2001||Mark A. Wolfe||Document retrieval system for retrieval of a first search document and a last search document from database|
|US6349275||24 Nov 1997||19 Feb 2002||International Business Machines Corporation||Multiple concurrent language support system for electronic catalogue using a concept based knowledge representation|
|US6356864 *||23 Jul 1998||12 Mar 2002||University Technology Corporation||Methods for analysis and evaluation of the semantic content of a writing based on vector length|
|US6389387 *||27 May 1999||14 May 2002||Sharp Kabushiki Kaisha||Method and apparatus for multi-language indexing|
|US6396951||23 Dec 1998||28 May 2002||Xerox Corporation||Document-based query data for information retrieval|
|US6415319||7 Feb 1997||2 Jul 2002||Sun Microsystems, Inc.||Intelligent network browser using incremental conceptual indexer|
|US6460036||5 Dec 1997||1 Oct 2002||Pinpoint Incorporated||System and method for providing customized electronic newspapers and target advertisements|
|US6466901 *||30 Nov 1998||15 Oct 2002||Apple Computer, Inc.||Multi-language document search and retrieval system|
|US6477488 *||10 Mar 2000||5 Nov 2002||Apple Computer, Inc.||Method for dynamic context scope selection in hybrid n-gram+LSA language modeling|
|US6493702||5 May 1999||10 Dec 2002||Xerox Corporation||System and method for searching and recommending documents in a collection using share bookmarks|
|US6510406 *||22 Mar 2000||21 Jan 2003||Mathsoft, Inc.||Inverse inference engine for high performance web search|
|US6602300 *||3 Sep 1998||5 Aug 2003||Fujitsu Limited||Apparatus and method for retrieving data from a document database|
|US6604103||9 Oct 2001||5 Aug 2003||Mark A. Wolfe||System and method for information retrieval employing a preloading procedure|
|US6611825||9 Jun 1999||26 Aug 2003||The Boeing Company||Method and system for text mining using multidimensional subspaces|
|US6615208 *||1 Sep 2000||2 Sep 2003||Telcordia Technologies, Inc.||Automatic recommendation of products using latent semantic indexing of content|
|US6654717||25 Feb 2002||25 Nov 2003||Apple Computer, Inc.||Multi-language document search and retrieval system|
|US6654740 *||8 May 2001||25 Nov 2003||Sunflare Co., Ltd.||Probabilistic information retrieval based on differential latent semantic space|
|US6697779||29 Sep 2000||24 Feb 2004||Apple Computer, Inc.||Combined dual spectral and temporal alignment method for user authentication by voice|
|US6701305 *||20 Oct 2000||2 Mar 2004||The Boeing Company||Methods, apparatus and computer program products for information retrieval and document classification utilizing a multidimensional subspace|
|US6751621 *||2 May 2000||15 Jun 2004||Manning & Napier Information Services, Llc.||Construction of trainable semantic vectors and clustering, classification, and searching using trainable semantic vectors|
|US6757646||25 Sep 2001||29 Jun 2004||Insightful Corporation||Extended functionality for an inverse inference engine based web search|
|US6778952 *||12 Sep 2002||17 Aug 2004||Apple Computer, Inc.||Method for dynamic context scope selection in hybrid N-gram+LSA language modeling|
|US6804662||27 Oct 2000||12 Oct 2004||Plumtree Software, Inc.||Method and apparatus for query and analysis|
|US6847966 *||24 Apr 2002||25 Jan 2005||Engenium Corporation||Method and system for optimally searching a document database using a representative semantic space|
|US6862710||22 Mar 2000||1 Mar 2005||Insightful Corporation||Internet navigation using soft hyperlinks|
|US6876998||13 Mar 2001||5 Apr 2005||Claritech Corporation||Method for cross-linguistic document retrieval|
|US6937986||28 Dec 2000||30 Aug 2005||Comverse, Inc.||Automatic dynamic speech recognition vocabulary based on external sources of information|
|US7024407||24 Aug 2001||4 Apr 2006||Content Analyst Company, Llc||Word sense disambiguation|
|US7031961||4 Dec 2000||18 Apr 2006||Google, Inc.||System and method for searching and recommending objects from a categorically organized information repository|
|US7051017||10 Dec 2002||23 May 2006||Insightful Corporation||Inverse inference engine for high performance web search|
|US7054856 *||29 Nov 2001||30 May 2006||Electronics And Telecommunications Research Institute||System for drawing patent map using technical field word and method therefor|
|US7062220||18 Apr 2001||13 Jun 2006||Intelligent Automation, Inc.||Automated, computer-based reading tutoring systems and methods|
|US7103594||1 Jul 2003||5 Sep 2006||Wolfe Mark A||System and method for information retrieval employing a preloading procedure|
|US7113943||5 Dec 2001||26 Sep 2006||Content Analyst Company, Llc||Method for document comparison and selection|
|US7124073||12 Feb 2002||17 Oct 2006||Sunflare Co., Ltd||Computer-assisted memory translation scheme based on template automaton and latent semantic index principle|
|US7137062||28 Dec 2001||14 Nov 2006||International Business Machines Corporation||System and method for hierarchical segmentation with latent semantic indexing in scale space|
|US7152065||1 May 2003||19 Dec 2006||Telcordia Technologies, Inc.||Information retrieval and text mining using distributed latent semantic indexing|
|US7174290||7 Jul 2003||6 Feb 2007||Apple Computer, Inc.||Multi-language document search and retrieval system|
|US7191118 *||12 Aug 2004||13 Mar 2007||Apple, Inc.||Method for dynamic context scope selection in hybrid N-gram+LSA language modeling|
|US7219073 *||2 Aug 2000||15 May 2007||Brandnamestores.Com||Method for extracting information utilizing a user-context-based search engine|
|US7257604||5 Aug 2003||14 Aug 2007||Wolfe Mark A||System and method for communicating information relating to a network resource|
|US7269598||26 May 2004||11 Sep 2007||Insightful Corporation||Extended functionality for an inverse inference engine based web search|
|US7283951||8 Nov 2001||16 Oct 2007||Insightful Corporation||Method and system for enhanced data searching|
|US7299247||14 Apr 2004||20 Nov 2007||Manning & Napier Information Services, Llc.||Construction of trainable semantic vectors and clustering, classification, and searching using trainable semantic vectors|
|US7302638||29 Aug 2003||27 Nov 2007||Wolfe Mark A||Efficiently displaying and researching information about the interrelationships between documents|
|US7328216 *||11 Aug 2003||5 Feb 2008||Recommind Inc.||System and method for personalized search, information filtering, and for generating recommendations utilizing statistical latent class models|
|US7346608||20 Sep 2004||18 Mar 2008||Bea Systems, Inc.||Method and apparatus for query and analysis|
|US7353164 *||13 Sep 2002||1 Apr 2008||Apple Inc.||Representation of orthography in a continuous vector space|
|US7369987||29 Dec 2006||6 May 2008||Apple Inc.||Multi-language document search and retrieval system|
|US7398201||19 Feb 2003||8 Jul 2008||Evri Inc.||Method and system for enhanced data searching|
|US7406456||14 Apr 2004||29 Jul 2008||Manning & Napier Information Services, Llc.||Construction of trainable semantic vectors and clustering, classification, and searching using trainable semantic vectors|
|US7415462||20 Jan 2006||19 Aug 2008||Content Analyst Company, Llc||Word sense disambiguation|
|US7421418||18 Feb 2004||2 Sep 2008||Nahava Inc.||Method and apparatus for fundamental operations on token sequences: computing similarity, extracting term values, and searching efficiently|
|US7428517 *||27 Feb 2003||23 Sep 2008||Brands Michael Rik Frans||Data integration and knowledge management solution|
|US7444356||14 Apr 2004||28 Oct 2008||Manning & Napier Information Services, Llc.|
|US7483871||1 Oct 2002||27 Jan 2009||Pinpoint Incorporated||Customized electronic newspapers and advertisements|
|US7483892 *||24 Jan 2005||27 Jan 2009||Kroll Ontrack, Inc.||Method and system for optimally searching a document database using a representative semantic space|
|US7523085||30 Sep 2005||21 Apr 2009||Buzzmetrics, Ltd An Israel Corporation||Topical sentiments in electronically stored communications|
|US7526425||13 Dec 2004||28 Apr 2009||Evri Inc.||Method and system for extending keyword searching to syntactically and semantically annotated data|
|US7558774||29 Oct 2007||7 Jul 2009||Nahava Inc.||Method and apparatus for fundamental operations on token sequences: computing similarity, extracting term values, and searching efficiently|
|US7562066||15 Nov 2001||14 Jul 2009||Hewlett-Packard Development Company, L.P.||Method of vector analysis for a document|
|US7580910||31 Mar 2006||25 Aug 2009||Content Analyst Company, Llc||Perturbing latent semantic indexing spaces|
|US7587307 *||18 Dec 2003||8 Sep 2009||Xerox Corporation||Method and apparatus for evaluating machine translation quality|
|US7596552||8 Sep 2006||29 Sep 2009||Buzzmetrics Ltd.||Method and system for extracting web data|
|US7600017||10 Jan 2007||6 Oct 2009||Buzzmetrics, Ltd.||System and method for scoring electronic messages|
|US7630986||27 Oct 2000||8 Dec 2009||Pinpoint, Incorporated||Secure data interchange|
|US7660783||31 Aug 2007||9 Feb 2010||Buzzmetrics, Inc.||System and method of ad-hoc analysis of data|
|US7702509||21 Nov 2006||20 Apr 2010||Apple Inc.||Unsupervised data-driven pronunciation modeling|
|US7716221 *||1 Jun 2007||11 May 2010||Behrens Clifford A||Concept based cross media indexing and retrieval of speech documents|
|US7720673||23 Feb 2007||18 May 2010||Apple Inc.||Method for dynamic context scope selection in hybrid N-GRAM+LSA language modeling|
|US7720792||7 Feb 2006||18 May 2010||Content Analyst Company, Llc||Automatic stop word identification and compensation|
|US7725414||16 Mar 2004||25 May 2010||Buzzmetrics, Ltd An Israel Corporation||Method for developing a classifier for classifying communications|
|US7756855||28 Aug 2007||13 Jul 2010||Collarity, Inc.||Search phrase refinement by search term replacement|
|US7765098||24 Apr 2006||27 Jul 2010||Content Analyst Company, Llc||Machine translation using vector space representations|
|US7779007||8 Jun 2007||17 Aug 2010||Ase Edge, Inc.||Identifying content of interest|
|US7818282 *||2 Jul 2004||19 Oct 2010||International Business Machines Corporation||System and method for the support of multilingual applications|
|US7844483||26 Feb 2007||30 Nov 2010||Buzzmetrics, Ltd.||System and method for predicting external events from electronic author activity|
|US7844484||26 Feb 2007||30 Nov 2010||Buzzmetrics, Ltd.||System and method for benchmarking electronic message activity|
|US7844566||11 May 2006||30 Nov 2010||Content Analyst Company, Llc||Latent semantic clustering|
|US7853600||20 May 2005||14 Dec 2010||Pinpoint, Incorporated||System and method for providing access to video programs and other data using customer profiles|
|US7860593||10 May 2007||28 Dec 2010||Canrig Drilling Technology Ltd.||Well prog execution facilitation system and method|
|US7877345||27 Feb 2009||25 Jan 2011||Buzzmetrics, Ltd.||Topical sentiments in electronically stored communications|
|US7903884 *||20 Jan 2005||8 Mar 2011||Bits Kabushikigaisha||System and computer software program for visibly processing an observed information's relationship with knowledge accumulations|
|US7912868||11 May 2005||22 Mar 2011||Textwise Llc||Advertisement placement method and system using semantic analysis|
|US7917488||3 Mar 2008||29 Mar 2011||Microsoft Corporation||Cross-lingual search re-ranking|
|US7953593||10 Mar 2009||31 May 2011||Evri, Inc.||Method and system for extending keyword searching to syntactically and semantically annotated data|
|US7983896 *||3 Mar 2005||19 Jul 2011||SDL Language Technology||In-context exact (ICE) matching|
|US8024331||2 Jun 2008||20 Sep 2011||Manning & Napier Information Services, Llc.|
|US8027876||14 Jun 2007||27 Sep 2011||Yoogli, Inc.||Online advertising valuation apparatus and method|
|US8041557 *||22 Feb 2006||18 Oct 2011||Fuji Xerox Co., Ltd.||Word translation device, translation method, and computer readable medium|
|US8041669||15 Dec 2010||18 Oct 2011||Buzzmetrics, Ltd.||Topical sentiments in electronically stored communications|
|US8051061||19 Feb 2008||1 Nov 2011||Microsoft Corporation||Cross-lingual query suggestion|
|US8131540||10 Mar 2009||6 Mar 2012||Evri, Inc.||Method and system for extending keyword searching to syntactically and semantically annotated data|
|US8156125||19 Feb 2008||10 Apr 2012||Oracle International Corporation||Method and apparatus for query and analysis|
|US8171026||16 Apr 2009||1 May 2012||Hewlett-Packard Development Company, L.P.||Method and vector analysis for a document|
|US8171032||4 Aug 2008||1 May 2012||Pinpoint, Incorporated||Providing customized electronic information|
|US8190608 *||30 Jun 2011||29 May 2012||Google Inc.||Systems and methods for using anchor text as parallel corpora for cross-language information retrieval|
|US8244739 *||14 Jan 2010||14 Aug 2012||Nec Corporation||Data retrieval device using a skip table|
|US8255401||28 Apr 2010||28 Aug 2012||International Business Machines Corporation||Computer information retrieval using latent semantic structure via sketches|
|US8271316||2 Feb 2006||18 Sep 2012||Buzzmetrics Ltd||Consumer to business data capturing system|
|US8326785||30 Sep 2008||4 Dec 2012||Microsoft Corporation||Joint ranking model for multilingual web search|
|US8346536||11 Nov 2008||1 Jan 2013||Eij Group Llc||System and method for multi-lingual information retrieval|
|US8347326||18 Dec 2007||1 Jan 2013||The Nielsen Company (US)||Identifying key media events and modeling causal relationships between key events and reported feelings|
|US8380719 *||18 Jun 2010||19 Feb 2013||Microsoft Corporation||Semantic content searching|
|US8386059||22 Dec 2010||26 Feb 2013||Canrig Drilling Technology Ltd.||Well prog execution facilitation system and method|
|US8429167||8 Aug 2006||23 Apr 2013||Google Inc.||User-context-based search engine|
|US8429184 *||14 Jun 2010||23 Apr 2013||Collarity Inc.||Generation of refinement terms for search queries|
|US8438178||25 Jun 2009||7 May 2013||Collarity Inc.||Interactions among online digital identities|
|US8442972||11 Oct 2007||14 May 2013||Collarity, Inc.||Negative associations for search results ranking and refinement|
|US8515811||29 Aug 2011||20 Aug 2013||Google Inc.||Online advertising valuation apparatus and method|
|US8521506||21 Sep 2006||27 Aug 2013||Sdl Plc||Computer-implemented method, computer software and apparatus for use in a translation system|
|US8583418||29 Sep 2008||12 Nov 2013||Apple Inc.||Systems and methods of detecting language and natural language strings for text to speech synthesis|
|US8594996||15 Oct 2008||26 Nov 2013||Evri Inc.||NLP-based entity recognition and disambiguation|
|US8600743||6 Jan 2010||3 Dec 2013||Apple Inc.||Noise profile determination for voice-related feature|
|US8614431||5 Nov 2009||24 Dec 2013||Apple Inc.||Automated response to and sensing of user activity in portable devices|
|US8620662||20 Nov 2007||31 Dec 2013||Apple Inc.||Context-aware unit selection|
|US8620793||1 Jun 2010||31 Dec 2013||Sdl International America Incorporated||Workflow management system|
|US8626763||4 Jun 2012||7 Jan 2014||Google Inc.||Server-side suggestion of preload operations|
|US8631010||18 May 2012||14 Jan 2014||Google Inc.||Systems and methods for using anchor text as parallel corpora for cross-language information retrieval|
|US8639694||23 May 2012||28 Jan 2014||Google Inc.||Client-side processing of preload operations|
|US8645125||30 Mar 2011||4 Feb 2014||Evri, Inc.||NLP-based systems and methods for providing quotations|
|US8645137||11 Jun 2007||4 Feb 2014||Apple Inc.||Fast, language-independent method for user authentication by voice|
|US8645372||29 Oct 2010||4 Feb 2014||Evri, Inc.||Keyword-based search engine results using enhanced query strategies|
|US8660849||21 Dec 2012||25 Feb 2014||Apple Inc.||Prioritizing selection criteria by automated assistant|
|US8670979||21 Dec 2012||11 Mar 2014||Apple Inc.||Active input elicitation by intelligent automated assistant|
|US8670985||13 Sep 2012||11 Mar 2014||Apple Inc.||Devices and methods for identifying a prompt corresponding to a voice input in a sequence of prompts|
|US8676904||2 Oct 2008||18 Mar 2014||Apple Inc.||Electronic devices with voice command and contextual data processing capabilities|
|US8677377||8 Sep 2006||18 Mar 2014||Apple Inc.||Method and apparatus for building an intelligent automated assistant|
|US8682649||12 Nov 2009||25 Mar 2014||Apple Inc.||Sentiment prediction from textual data|
|US8682667||25 Feb 2010||25 Mar 2014||Apple Inc.||User profiling for selecting user specific voice input processing information|
|US8688446||18 Nov 2011||1 Apr 2014||Apple Inc.||Providing text input using speech data and non-speech data|
|US8700604||16 Oct 2008||15 Apr 2014||Evri, Inc.||NLP-based content recommender|
|US8706472||11 Aug 2011||22 Apr 2014||Apple Inc.||Method for disambiguating multiple readings in language conversion|
|US8706503||21 Dec 2012||22 Apr 2014||Apple Inc.||Intent deduction based on previous user interactions with voice assistant|
|US8712776||29 Sep 2008||29 Apr 2014||Apple Inc.||Systems and methods for selective text to speech synthesis|
|US8713021||7 Jul 2010||29 Apr 2014||Apple Inc.||Unsupervised document clustering using latent semantic density analysis|
|US8713119||13 Sep 2012||29 Apr 2014||Apple Inc.||Electronic devices with voice command and contextual data processing capabilities|
|US8718047||28 Dec 2012||6 May 2014||Apple Inc.||Text to speech conversion of text messages from mobile communication devices|
|US8718802||5 Feb 2013||6 May 2014||Canrig Drilling Technology Ltd.||Well prog execution facilitation system and method|
|US8719006||27 Aug 2010||6 May 2014||Apple Inc.||Combined statistical and rule-based part-of-speech tagging for text-to-speech synthesis|
|US8719014||27 Sep 2010||6 May 2014||Apple Inc.||Electronic device with text error correction based on voice recognition data|
|US8725739||1 Nov 2011||13 May 2014||Evri, Inc.||Category-based content recommendation|
|US8731942||4 Mar 2013||20 May 2014||Apple Inc.||Maintaining context information between user interactions with a voice assistant|
|US8750468||5 Oct 2010||10 Jun 2014||Callspace, Inc.||Contextualized telephony message management|
|US8751238||15 Feb 2013||10 Jun 2014||Apple Inc.||Systems and methods for determining the language to use for speech generated by a text to speech engine|
|US8762156||28 Sep 2011||24 Jun 2014||Apple Inc.||Speech recognition repair using contextual information|
|US8762469||5 Sep 2012||24 Jun 2014||Apple Inc.||Electronic devices with voice command and contextual data processing capabilities|
|US8768702||5 Sep 2008||1 Jul 2014||Apple Inc.||Multi-tiered voice feedback in an electronic device|
|US8775442||15 May 2012||8 Jul 2014||Apple Inc.||Semantic search using a single-source semantic model|
|US8781836||22 Feb 2011||15 Jul 2014||Apple Inc.||Hearing assistance system for providing consistent human speech|
|US8793715||20 Nov 2012||29 Jul 2014||The Nielsen Company (Us), Llc||Identifying key media events and modeling causal relationships between key events and reported feelings|
|US8799000||21 Dec 2012||5 Aug 2014||Apple Inc.||Disambiguation based on active input elicitation by intelligent automated assistant|
|US8799461||23 Mar 2012||5 Aug 2014||Apple Inc.||System for collecting, analyzing, and transmitting information relevant to transportation networks|
|US8805676||25 Sep 2012||12 Aug 2014||Abbyy Infopoisk Llc||Deep model statistics method for machine translation|
|US8805842||30 Mar 2012||12 Aug 2014||Her Majesty The Queen In Right Of Canada, As Represented By The Minister Of National Defence, Ottawa||Method for displaying search results|
|US8812294||21 Jun 2011||19 Aug 2014||Apple Inc.||Translating phrases from one language into another using an order-based set of declarative rules|
|US8812541||12 Mar 2013||19 Aug 2014||Collarity, Inc.||Generation of refinement terms for search queries|
|US8838633||11 Aug 2011||16 Sep 2014||Vcvc Iii Llc||NLP-based sentiment analysis|
|US8856096||16 Nov 2006||7 Oct 2014||Vcvc Iii Llc||Extending keyword searching to syntactically and semantically annotated data|
|US8856156||5 Oct 2012||7 Oct 2014||Cerner Innovation, Inc.||Ontology mapper|
|US8862252||30 Jan 2009||14 Oct 2014||Apple Inc.||Audio user interface for displayless electronic device|
|US8862591 *||22 Aug 2007||14 Oct 2014||Twitter, Inc.||System and method for evaluating sentiment|
|US8862595 *||18 Nov 2011||14 Oct 2014||Google Inc.||Language selection for information retrieval|
|US8874427 *||1 Jul 2011||28 Oct 2014||Sdl Enterprise Technologies, Inc.||In-context exact (ICE) matching|
|US8874727||13 Jul 2010||28 Oct 2014||The Nielsen Company (Us), Llc||Methods, apparatus, and articles of manufacture to rank users in an online social network|
|US8875038||19 Jan 2011||28 Oct 2014||Collarity, Inc.||Anchoring for content synchronization|
|US8892418||20 Jul 2012||18 Nov 2014||Abbyy Infopoisk Llc||Translating sentences between languages|
|US8898568||9 Sep 2008||25 Nov 2014||Apple Inc.||Audio user interface|
|US8903810||16 Oct 2008||2 Dec 2014||Collarity, Inc.||Techniques for ranking search results|
|US8904273||2 Jul 2004||2 Dec 2014||International Business Machines Corporation||System and method of format specification|
|US8918309||20 Dec 2012||23 Dec 2014||Abbyy Infopoisk Llc||Deep model statistics method for machine translation|
|US8935148||14 Dec 2009||13 Jan 2015||Sdl Plc||Computer-assisted natural language translation|
|US8935150||5 Sep 2013||13 Jan 2015||Sdl Plc||Dynamic generation of auto-suggest dictionary for natural language translation|
|US8935167||25 Sep 2012||13 Jan 2015||Apple Inc.||Exemplar-based latent perceptual modeling for automatic speech recognition|
|US8954469||14 Mar 2008||10 Feb 2015||Vcvciii Llc||Query templates and labeled search tip system, methods, and techniques|
|US8959011||26 Jan 2012||17 Feb 2015||Abbyy Infopoisk Llc||Indicating and correcting errors in machine translation systems|
|US8971630||19 Jun 2012||3 Mar 2015||Abbyy Development Llc||Fast CJK character recognition|
|US8977255||3 Apr 2007||10 Mar 2015||Apple Inc.||Method and system for operating a multi-function portable electronic device using voice-activation|
|US8989485||14 Oct 2013||24 Mar 2015||Abbyy Development Llc||Detecting a junction in a text line of CJK characters|
|US8996376||5 Apr 2008||31 Mar 2015||Apple Inc.||Intelligent text-to-speech conversion|
|US9053089||2 Oct 2007||9 Jun 2015||Apple Inc.||Part-of-speech tagging using latent analogy|
|US9053090||20 Jun 2012||9 Jun 2015||Abbyy Infopoisk Llc||Translating texts between languages|
|US9069750 *||30 Jun 2011||30 Jun 2015||Abbyy Infopoisk Llc||Method and system for semantic searching of natural language texts|
|US9075783||22 Jul 2013||7 Jul 2015||Apple Inc.||Electronic device with text error correction based on voice recognition data|
|US9075864||31 Dec 2010||7 Jul 2015||Abbyy Infopoisk Llc||Method and system for semantic searching using syntactic and semantic analysis|
|US9092416||31 Jan 2014||28 Jul 2015||Vcvc Iii Llc||NLP-based systems and methods for providing quotations|
|US9098489 *||30 Jun 2011||4 Aug 2015||Abbyy Infopoisk Llc||Method and system for semantic searching|
|US20020016786 *||4 Dec 2000||7 Feb 2002||Pitkow James B.||System and method for searching and recommending objects from a categorically organized information repository|
|US20040068396 *||22 Sep 2003||8 Apr 2004||Takahiko Kawatani||Method of vector analysis for a document|
|US20040133574 *||7 Jan 2003||8 Jul 2004||Science Applications International Corporaton||Vector space method for secure information sharing|
|US20040162827 *||18 Feb 2004||19 Aug 2004||Nahava Inc.||Method and apparatus for fundamental operations on token sequences: computing similarity, extracting term values, and searching efficiently|
|US20040193414 *||14 Apr 2004||30 Sep 2004||Manning & Napier Information Services, Llc|
|US20040199505 *||14 Apr 2004||7 Oct 2004||Manning & Napier Information Services, Llc|
|US20040199546 *||14 Apr 2004||7 Oct 2004||Manning & Napier Information Services, Llc|
|US20040220944 *||1 May 2003||4 Nov 2004||Behrens Clifford A||Information retrieval and text mining using distributed latent semantic indexing|
|US20040221235 *||8 Nov 2001||4 Nov 2004||Insightful Corporation||Method and system for enhanced data searching|
|US20050015239 *||12 Aug 2004||20 Jan 2005||Bellegarda Jerome R.||Method for dynamic context scope selection in hybrid N-gramlanguage modeling|
|US20050021517 *||26 May 2004||27 Jan 2005||Insightful Corporation||Extended functionality for an inverse inference engine based web search|
|US20050038866 *||21 Oct 2002||17 Feb 2005||Sumio Noguchi||Information search support apparatus, computer program, medium containing the program|
|US20050097092 *||20 Sep 2004||5 May 2005||Ripfire, Inc., A Corporation Of The State Of Delaware||Method and apparatus for query and analysis|
|US20050137854 *||18 Dec 2003||23 Jun 2005||Xerox Corporation||Method and apparatus for evaluating machine translation quality|
|US20050197827 *||3 Mar 2005||8 Sep 2005||Russ Ross||In-context exact (ICE) matching|
|US20050203889 *||20 Jan 2005||15 Sep 2005||Okubo, Kousaku.||System and computer software program for visibly processing an observed information's relationship with knowledge accumulations|
|US20050216516 *||11 May 2005||29 Sep 2005||Textwise Llc||Advertisement placement method and system using semantic analysis|
|US20050251514 *||13 Apr 2005||10 Nov 2005||International Business Machines Corporation||Evaluation and cluster formation based on element correlation|
|US20050267871 *||13 Dec 2004||1 Dec 2005||Insightful Corporation||Method and system for extending keyword searching to syntactically and semantically annotated data|
|US20060004738 *||2 Jul 2004||5 Jan 2006||Blackwell Richard F||System and method for the support of multilingual applications|
|US20060005112 *||2 Jul 2004||5 Jan 2006||David Lilly||System and method of report layout|
|US20060005127 *||2 Jul 2004||5 Jan 2006||Ferguson Kevin M||System and method of format specification|
|US20080154883 *||22 Aug 2007||26 Jun 2008||Abdur Chowdhury||System and method for evaluating sentiment|
|US20090077126 *||18 Sep 2008||19 Mar 2009||Nec (China) Co,. Ltd||Method and system for calculating competitiveness metric between objects|
|US20110035403 *||14 Jun 2010||10 Feb 2011||Emil Ismalon||Generation of refinement terms for search queries|
|US20110099177 *||14 Jan 2010||28 Apr 2011||Nec Corporation||Data retrieval device|
|US20110257963 *||20 Oct 2011||Konstantin Zuev||Method and system for semantic searching|
|US20110258196 *||30 Dec 2008||20 Oct 2011||Skjalg Lepsoy||Method and system of content recommendation|
|US20110270607 *||3 Nov 2011||Konstantin Zuev||Method and system for semantic searching of natural language texts|
|US20110314024 *||18 Jun 2010||22 Dec 2011||Microsoft Corporation||Semantic content searching|
|US20120095747 *||1 Jul 2011||19 Apr 2012||Russ Ross||In-context exact (ice) matching|
|US20120309363 *||30 Sep 2011||6 Dec 2012||Apple Inc.||Triggering notifications associated with tasks items that represent tasks to perform|
|USRE36727 *||4 Apr 1997||6 Jun 2000||Kageneck; Karl-Erbo G.||Method of indexing and retrieval of electronically-stored documents|
|CN101512521B||1 Jun 2007||16 Jan 2013||Tti发明有限责任公司||Concept based cross media indexing and retrieval of speech documents|
|EP0750266A1 *||29 Apr 1996||27 Dec 1996||Sharp Kabushiki Kaisha||Document classification unit and document retrieval unit|
|EP0964344A2 *||28 May 1999||15 Dec 1999||Sharp Kabushiki Kaisha||Method of and apparatus for forming an index, use of an index and a storage medium|
|EP1323067A1 *||25 Sep 2001||2 Jul 2003||Insightful Corporation||Extended functionality for an inverse inference engine based web search|
|EP1909203A2||2 Dec 2004||9 Apr 2008||Sony Corporation||Apparatus, method and program for information processing|
|WO1995000896A2 *||13 Jun 1994||5 Jan 1995||Libertech Inc||Method and apparatus for indexing searching and displaying data|
|WO1997008604A2 *||14 Aug 1996||6 Mar 1997||Univ Syracuse||Multilingual document retrieval system and method using semantic vector matching|
|WO2002021335A1 *||17 Aug 2001||14 Mar 2002||Telcordia Tech Inc||Automatic recommendation of products using latent semantic indexing of content|
|WO2002027536A1 *||25 Sep 2001||4 Apr 2002||Insightful Corp||Extended functionality for an inverse inference engine based web search|
|WO2002041557A2 *||15 Nov 2001||23 May 2002||Hewlett Packard Co||Method of vector analysis for a document|
|WO2003069503A1 *||30 Jan 2003||21 Aug 2003||Idea Inc||New computer-assisted memory translation scheme based on template automaton and latent semantic index principle|
|U.S. Classification||704/9, 704/2, 707/E17.061|
|International Classification||G06F17/27, G06F17/28, G06F17/30|
|Cooperative Classification||G06F17/289, G06F17/274, G06F17/30634|
|European Classification||G06F17/27G, G06F17/30T2, G06F17/28U|
|17 Jul 1991||AS||Assignment|
Owner name: BELL COMMUNICATIONS RESEARCH, INC.,, NEW JERSEY
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST.;ASSIGNORS:LANDAUER, THOMAS K.;LITTMAN, MICHAEL L.;REEL/FRAME:005786/0557
Effective date: 19910716
|30 Sep 1997||FPAY||Fee payment|
Year of fee payment: 4
|4 Oct 1999||AS||Assignment|
|25 Sep 2001||FPAY||Fee payment|
Year of fee payment: 8
|18 Apr 2005||AS||Assignment|
|7 Jun 2005||AS||Assignment|
Owner name: CONTENT ANALYST COMPANY LLC, VIRGINIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SCIENCE APPLICATIONS INTERNATIONAL CORPORATION;REEL/FRAME:016097/0468
Effective date: 20050606
|6 Sep 2005||FPAY||Fee payment|
Year of fee payment: 12